Optimized, Automated Striatal Uptake AnalysisApplied to SPECT Brain Scans of Parkinson’sDisease Patients

I. George Zubal, Michele Early, Olive Yuan, Danna Jennings, Kenneth Marek, and John P. Seibyl

Institute for Neurodegenerative Disorders, and Molecular Neuroimaging, LLC, New Haven, Connecticut

Reliable quantitative dopamine transporter imaging is critical forearly and accurate diagnosis of Parkinson’s disease (PD). Imagequantitation is made difficult by the variability introduced bymanual interventions during the quantitative processing steps.A fully automated objective striatal analysis (OSA) programwas applied to dopamine transporter images acquired from PDsubjects with early symptoms of suspected parkinsonism andcompared with manual analysis by a trained image-processingtechnologist. Methods: A total of 101 123I-b-CIT SPECT scanswere obtained of subjects recruited to participate in the Query-PD Study. Data were reconstructed and then analyzed accordingto a package of scripts (OSA) that reorients the SPECT brain vol-ume to the standard geometry of an average scan, automaticallylocates the striata and occipital structures, locates the caudateand putamen, and calculates the background-subtracted striataluptake ratio (V3$). The striatal uptake ratio calculated by OSA wascompared with manual analysis by a trained image-processingtechnologist. Several parameters were varied in the automatedanalysis, including the number of summed transverse slicesand the size and separation of the regions of interest applied tothe caudate and putamen to determine the optimum OSA analy-sis. The parameters giving V3$ with the closest correlation to themanual analysis were accepted as optimal. Results: The optimalcomparison between the V3$ obtained by the human analyst andthat obtained by the automated OSA analysis yielded a correla-tion coefficient of 0.96. Conclusion: Our optimized OSA deliversV3$ evaluations that closely correlate with a similar evaluationmanually applied by a highly trained image-processing technol-ogist.

Key Words: PACS; neurology; SPECT; Parkinson’s disease;analysis; automated

J Nucl Med 2007; 48:857–864DOI: 10.2967/jnumed.106.037432

Parkinson’s disease (PD) is characterized by the slowlyprogressive degeneration of nigrostriatal dopaminergicneurons as shown in Figure 1. Identification of the loss ofdopaminergic neurons has led to effective symptomatic

therapy using dopamine replacement but does not preventprogressive neuronal loss and disability (1).

More accurate and objective assessment of disease onsetand disease progression in individuals with PD is becomingincreasingly important as neuroprotective and neurorestor-ative strategies for PD are developed and tested (2,3). Theseneuroprotective strategies further underscore the need forearly and accurate diagnosis because typically about 50%of dopaminergic neurons have been lost by the time ofsymptom onset.

Although clinical diagnosis correlates well with patho-logic diagnosis after long-term clinical follow-up (4), earlyin the course of PD the diagnoses may be difficult and ofteninaccurate. The most commonly mistaken diagnoses for PDinclude essential tremor, vascular parkinsonism, drug-inducedparkinsonism, and Alzheimer’s disease (5,6). In addition,subtle symptoms of motor slowness easily confused withparkinsonism may be found in 32%–35% of elderly sub-jects, often complicating diagnosis in this most common agegroup for PD (7,8). More recently, the need for diagnosis toenable therapy even before symptoms are manifested, dur-ing the premotor period of PD, has further highlighted theneed for objective, accurate early diagnostic tools (9).

During the past decade, in vivo neuroimaging targetingthe dopamine transporter (DAT) has emerged as a reliablediagnostic tool for PD and related parkinsonian syndromes.Numerous studies have demonstrated that DAT ligands andSPECT or PET identify individuals with PD and distinguishthem from healthy subjects or individuals with movementdisorders not associated with dopamine depletion, such asessential tremor, nonspecific gait disorders, drug-inducedparkinsonism, and psychogenic parkinsonism (10–14). Morerecently, studies have further demonstrated that DAT imag-ing adds to diagnostic accuracy in patients with suspectedPD at the earliest stages of illness (15,16).

Reliable, rapid, easy-to-administer automated tools forquantitative analysis of DAT images would potentially im-prove diagnostic accuracy and enable more effective, wide-spread use of DAT imaging technology in clinical practice.Other investigators have developed methods to quantitatethe uptake level of DATs by analyzing the associated SPECT

Received Oct. 20, 2006; revision accepted Mar. 6, 2007.For correspondence contact: I. George Zubal, PhD, Institute for

Neurodegenerative Disorders, 60 Temple St., Suite 8, New Haven, CT 06510.E-mail: gzubal@indd.orgCOPYRIGHT ª 2007 by the Society of Nuclear Medicine, Inc.

AUTOMATED PARKINSON’S ANALYSIS • Zubal et al. 857

scans (17–21). In this study, we tested an automated DAT-imaging objective striatal analysis (OSA) program specif-ically for 123I-b-CIT in a large clinical cohort and comparedthe results with a well-established manual analysis by a trainedimage-processing technologist.

MATERIALS AND METHODS

The clinical cohort evaluated in this study was recruited as partof the Query-PD Study (15), a study to determine the diagnosticaccuracy of DAT imaging in subjects identified by communityneurologists as suspected of having, but not yet diagnosed with,PD or parkinsonian syndrome. All subjects signed informed con-sent and were clinically evaluated at Molecular NeuroImaging,LLC, by a movement disorders expert. Subjects met the followinginclusion criteria: age greater than 21 y, any parkinsonian or extra-pyramidal symptoms as determined by the primary neurologist,parkinsonian symptoms for less than 2 y, and no significant abnor-malities on screening laboratory studies, including complete bloodcell count, Chem-20, and urinalysis.

All subjects underwent 123I-b-CIT DAT imaging at MolecularNeuroImaging, LLC, using a previously described and operation-alized image acquisition (16). All patients received supersaturatedpotassium iodide solution (800 mg orally) before tracer injection.After injection of 370 6 57 MBq (10.0 6 1.6 mCi), projectiondata were obtained in a 128 · 128 matrix on a Picker Prism 3-headcamera (Philips Medical Systems) mounted with fanbeam collima-tors, by acquiring 120 angles over 40 min. The data were recon-structed using the manufacturer-supplied (Odyssey VP 8.5A5)standard filtered backprojection algorithm and applying a post-reconstruction 3-dimensional Butterworth smoothing filter (order,6; cutoff, 0.24 cycles per pixel). With the radius of rotation set to

15 cm, this process results in tomographic resolutions of 2.1 mmin the x- and y-directions and 3.56 mm in the z-direction. Atten-uation was corrected using the Chang zero-order correction (22),applying a linear attenuation coefficient of m 5 0.11 cm21. Theneuroimaging personnel were unaware of any clinical information.

For quantitation of DAT SPECT data, stringent techniquesinvolving a highly operationalized series of steps were performedby an experienced image processor. These manual steps, usingthe manufacturer-supplied processing software, included, first,2-dimensional Butterworth filtration (order, 10; cutoff, 0.24) of the120 projection images acquired for each scan; second, reconstruc-tion using filtered backprojection and a simple ramp filter; third,manual reorientation of the reconstructed imaging volume along thecanthomeatal line, determined from external radioactivity-containingfiducial markers placed on the subject’s skin; fourth, identificationof axial slices with striatal activity, summation of the specifiednumber of slices, and application of a Chang attenuation correction;fifth, placement of a regional template on the left and right caudateand putamen and the occipital background region; and sixth,extraction of count density data for determination of regional striatalV3$ as indicated in the following equation: V3$ 5 (regional striatalcount density 2 occipital count density)/occipital count density.

After completion of the predefined reconstruction, manualimage processing included several steps. The canthomeatal lineand the mid plane were visually determined between the 2hemispheres so as to rotate them into a horizontal and a verticalorientation, respectively. Once reoriented, 8 transverse slices werevisually selected as representing the centralmost slices containingthe striata. These selected slices were added together to create asingle slice for region-of-interest (ROI) placement. The image-processing platform has a predefined template containing theattenuation correction ellipse and associated ROIs correspondingto the expected position of the caudate and putamen. After theattenuation ellipse was placed onto the transverse slice, the striatalROIs could be manually adjusted until the operator decided thatthey were optimally overlaid onto the caudate and putamen.Similarly, a large semicircular ROI could be repositioned untilthe operator decided that it best measured the distribution ofcounts found in the occipital lobes. Once these ROIs were finallypositioned, V3$ was calculated. The positioning of the attenuationellipse and ROIs over the striatal and occipital structures wasbased on uptake intensities seen on the transverse images of the123I-b-CIT SPECT scan. Because no accompanying anatomicscans (CT or MRI) are routinely acquired for subjects undergoing123I-b-CIT SPECT, the operator relied on functional informationalone when placing these ROIs. Fortunately, even for patients withadvanced disease, the caudate demonstrated a relatively high ratioof tracer uptake, compared with the surrounding nonspecific back-ground, and could easily be identified.

Automated Image ProcessingA noninteractive OSA software package was developed for

application to the already-reconstructed and attenuation-correctedSPECT brain volumes and accomplished the third, fourth, fifth,and sixth steps above. These Matlab scripts run under Matlabversion 6.5.1 (The MathWorks, Inc.) and perform the image-processing steps outlined below.

Reorientation of the Reconstructed Imaging VolumeThe patient’s SPECT scan was reoriented by adoption of a

stylized average scan, which represented an averaged brain volume

FIGURE 1. DAT imaging in PD patient followed for 4 y. Visually,there is decreased uptake of radiotracer in striatum, indicative ofongoing loss of presynaptic dopaminergic function. Although theloss is easily noted, it is difficult to determine clinical meaning-fulness of this extent of signal change without quantitation.

858 THE JOURNAL OF NUCLEAR MEDICINE • Vol. 48 • No. 6 • June 2007

of 26 randomly selected subjects (13 healthy and 13 with PD).This average scan was manually reoriented into a position inwhich the canthomeatal line lay parallel to the horizontal planeswithin the imaging matrix. Once this standard orientation wasavailable and stored, all other subjects’ scans were geometricallyregistered to this average scan using an implementation of themutual information algorithm (23). Figure 2 shows the displayscreen for this registration process. The left half of this figure showsthe stylized average scan, in which a sum of all sagittal, transverse,and coronal slices is displayed in the top left, bottom left, and bottomright panels, respectively. The right half of this figure shows asimilarly displayed example of the subject’s summed sagittal,transverse, and coronal views. The green, blue, and red contoursoverlaid onto the subject’s images correspond to the isocontour linestransferred from the average scan and help to show the accuracy ofthe registration.

The rationale for registering the patient’s scan to the averagescan is to ensure that the patient’s brain volume is oriented in astandard position, that is, to correct for any rotation around the x-,y-, or z-axis of the patient’s head within the imaging camera. Thiscorrection is important because the subsequent location of striataand placement of ROIs (striatal and background) can be influ-enced by variations in the pitch, yaw, and roll orientations of thehead. However, to optimize registration between the patient’s scanand the average scan (despite variation in individual head size),one needs to apply a scaling. Hence, the average scan is scaled(but not rotated or translated) independently in the x-, y-, andz-axes when the patient’s scans are registered to the average scan,yielding an optimal registration of the patient’s scan with theorientation of the average scan, without scaling or warping thepatient’s scan (which could potentially introduce additional inter-polation errors later in the quantitation).

Identification of StriataIdentification of the striata is fairly simple. Even in advanced

disease, the target-to-background ratio is high and the striata aresmall in relation to the whole brain. The 2-step process of locatingthe centers of the 2 striata within the patient’s brain volume isshown in Figure 3.

The location of the horizontal central slice passing through thecenter of the striata is shown in the left half of Figure 3. Bycalculating the top-to-bottom profile of the summed intensities inthe sagittal view, one can easily identify the location of the centralslice through the striata as the maximum along this profile. Thesmall white panel shows the graph of this summed profile, and the

green arrow points along the central slice to the location of thestriata. The red lines above and below show the selection of Ntransverse slices for subsequent processing (in this example, N 5

9). The right half of Figure 3 shows the same subject’s brainvolume when the intensities above and below the selected N slicesare set to zero. Maxima along the summed profiles of the sagittaland coronal images localize the position of the 2 striata. The greenarrows point back from the maxima along the profiles to thelocation of the 2 striata.

All subsequent processing from this point on is applied to theseselected N (summed) transverse slices. Selecting the optimal valuefor N is described in the section on parameter optimization.

Measurement of Head Size and Placement of ROIsTo properly place the ROIs on the caudate, putamen, and

occipital lobes, one must calibrate each subject’s relative headsize. The calibration method is shown in the left half of Figure 4,in which a summed profile is calculated along the anterior-to-posterior direction of the N selected slices and is displayed in thewhite panel.

The peak corresponding to the striata appears superimposed ona fairly well defined plateau of nonspecific activity distributedthroughout the brain. The edge of the occipital lobe is taken as the50% intensity threshold found on the posterior portion of thisplateau profile. The location of the striata and edge of the occipitallobe are shown with red lines drawn on the sagittal and transverseimage displays and on the profile in the white panel. The pixeldistance between these red lines is the head size.

The head size of the average scan is 55 pixels. When subjectsare analyzed, the size (s) and separation (d) of the ROIs to beapplied on the caudate and putamen, as well as the size of theoccipital lobe ROI, are linearly scaled in proportion to eachsubject’s head size.

Once the scaling for the ROIs is determined, as shown in theright half of Figure 4, a larger caudate ROI (green circle) is movedwithin the N-summed slices image until the ROI yields a maxi-mum value of total counts; this is done for both the left and theright striatum. The small vertical and horizontal arrows depict thissearching algorithm. Within each green circle, a smaller caudateROI (red circle) is placed at its center and is used for the countdensity measurement of each caudate. Given a default separationof the caudate to putamen, d, the putamen ROI (yellow circle)orbits around the caudate ROI at the appropriately scaled distanceuntil the total counts within the ROI yield a maximum; the curvedyellow arrows depict this orbiting, searching algorithm.

FIGURE 2. Display showing examplepatient’s brain scan (right half of figure)being registered to standard averagescan (left half of figure).

AUTOMATED PARKINSON’S ANALYSIS • Zubal et al. 859

The default occipital ROI (large red truncated annulus) isscaled equally in the x- and y-directions and is placed just insidethe edge of the occipital lobe.

The regional striatal count density is calculated as the averageof the 2 caudate and 2 putamen ROI count densities, the occipitalcount density is calculated from the occipital ROI, and V3$ iscalculated according to Equation 1.

Optimizing the Parameters for AnalysisThe gold standard for testing the accuracy of the automated

OSA program is the V3$ manually obtained by a trained, highlyexperienced image-processing technologist. However, it was notknown which values of N, s, and d would correlate best with thehuman analysis. To determine the optimal values for N, s, and d,all 101 patients were batch analyzed, by which the parameter Ntook on values of 4, 5, 7, 9, 11, 13, 15, and 17 (slices); the parameter stook on values of 1, 2, and 3 (pixels); and the parameter d took onvalues of 8, 9, 10, 11, and 12 (pixels); that is, all 101 patients wereanalyzed 120 times (each time with a different set of N, s, and dparameters) and then compared with manual analysis by calculationof the linear regression slope and its associated R value.

RESULTS

Figure 5 shows a portion of the 120 evaluations performedusing OSA. In this section of the analysis, the parameters N,

s, and d were varied between 7 and 13, between 1 and 4, and

between 9 and 12, respectively (plotted along the horizontalaxis). For each set of these parameters, 101 V3$ evaluationswere compared with the manual analysis, yielding an R value(plotted along the vertical axis). The cyclic oscillations of theR values demonstrated that as the values of N, s, and d werevaried, the correlation between the manual and automateddeterminations of V3$ for this group of subjects changed.

The accuracy of the OSA analysis was optimal when N,s, and d took on values of 9, 1, and 8, respectively. The

scatter plot for this case is shown in Figure 6. The inter-pretation of s 5 1 is the use of a 3 · 3 pixel matrix ROI

(i.e., a central pixel 6 1 neighbor pixel in each direction).The precision of the OSA analysis was optimal when N,

s, and d took on values of 13, 0, and 10, respectively,

corresponding to the maximum R value in Figure 5. Thescatter plot for this case is shown in Figure 7. The inter-

pretation of s 5 0 is the use of a single-pixel ROI.Accuracy (measured by the linear regression slope) and

precision (measured by the R value) changed as the com-

bination of parameters N, s, and d took on various values.Figure 8 shows this variation when the regression slope and

R value were plotted as a function of N, s, and d. Here, N, s,

FIGURE 3. Display showing selectionof horizontal striatal midplane (left half offigure) and subsequent localization of leftand right striatum (right half of figure),both based on location of maxima ineither sagittal or coronal profiles.

FIGURE 4. Display showing automaticmeasurement of head size (left half offigure), whereby distance from midpointbetween the 2 striata to edge of occipitallobes is used to scale subsequent ROIplacement. Right half of figure shows place-ment of caudate ROIs (small red circles),putamen ROIs (small yellow circles), andoccipital ROI (large truncated annulus).

860 THE JOURNAL OF NUCLEAR MEDICINE • Vol. 48 • No. 6 • June 2007

and d vary from values of 7, 0, and 8, respectively, at theleft side of the horizontal axis to values of 17, 1, and 9,respectively, at the right side of the axis. Because optimalslope and R value would be equal to 1.0, it is interesting tonote that the values of R and the slope cross each other at apoint corresponding to N, s, and d values of 9, 1, and 8,respectively (the parameters that yielded the most accurateresults), and that the plot of the slope contains 6 repeatingcycles that correspond to the selection of increasing Nvalues. Increasing the N values had the strongest influenceon decreasing the accuracy of OSA.

DISCUSSION

Our optimized, automated analysis (OSA) delivered 123I-b-CIT striatal V3$ values for 101 patients with suspectedparkinsonism that strongly correlated (r 5 0.96) with the sameevaluations conducted by a highly trained image-processingresearcher. OSA provides a rapid, efficient, flexible, andreliable tool for quantitative DAT imaging. One particularadvantage of OSA is that the program can be exported to

FIGURE 5. Graph of R value versus combination of N, S, andd parameters used by OSA. R value comes from comparison ofOSA V3$ values to those obtained by experienced imageprocessing researcher. Values of N vary from 9 to 15 from left toright along axis; values of s change from 1 to 4 and cycle moreslowly from left to right than do values of d, which cycles morequickly through values of 8 to 12.

FIGURE 6. Scatter plot of OSA V3$ evaluations comparedwith those of experienced researcher for OSA parameters N, s,and d set to 9, 1, and 8, respectively. These parameters yieldedoptimal slope of 0.93 and demonstrated best accuracy for OSAanalysis.

FIGURE 7. Scatter plot of OSA V3$ evaluations comparedwith those of experienced researcher for OSA parameters N, s,and d set to 13, 0, and 10, respectively. These parametersyielded optimal R value of 0.96 (R2 5 0.93) and demonstratedbest precision for OSA analysis.

FIGURE 8. Plot of both R value and regression slope vs.combination of N, S, and d parameters used by OSA. Pa-rameters along horizontal axis are similar to those in Figure 5 inthat, here, values of N vary from 3 to 17 from left to right alongaxis; values of s change from 1 to 4 and cycle more slowly fromleft to right than do values of d, which cycles more quicklythrough values of 8 to 12.

AUTOMATED PARKINSON’S ANALYSIS • Zubal et al. 861

multiple imaging sites, where the identical analysis is con-ducted as it would be by a central image-processing labo-ratory. The availability of an automated imaging-analysisprogram is critical to the widespread use of DAT imaging,particularly at sites with little experience in analyzing DATSPECT images or DAT radiotracers involving multiple dy-namic acquisitions after radiotracer injection.

DAT imaging for PD diagnosis is commercially availablein most European countries and is currently under devel-opment in the United States. Accurate and reliable quan-titation of DAT imaging outcomes is critical to early diagnosis,and the need for earlier PD diagnosis to allow earliertreatment is growing. It is well known that PD has aprolonged presymptomatic period during which neurode-generation occurs but symptoms have not yet manifested.Imaging offers the best opportunity to accurately push backthe diagnosis to the presymptomatic period (24). Althoughimaging is unlikely to be used as a primary screen for PDbecause of cost, coupling of DAT imaging with more widelyavailable PD biomarkers may effectively identify subjectswho might be presymptomatic (25). As DAT imaging is usedin increasing numbers of subjects and imaging sites, OSAwould provide both a reliable and a practical method toensure accurate quantitation of imaging outcomes.

OSA is fully automated, that is, after the file names forthe attenuation-corrected reconstructed brain volumes havebeen selected, all subsequent processing is accomplishedwithout any manual intervention and is completed within2 min (running on Windows XP [Microsoft] with a 1-GHzprocessor), 90% of which time is needed for 3-dimensionalregistration to the average scan. For the clinical cohort inthis study, using a simple batch script, 101 subjects werereanalyzed within 3.5 h. This brief processing time allowsfurther analyses for evaluating the effect of attenuationcorrection, scatter correction, iterative or filtered backpro-jection reconstruction parameters, and other variables onthe resulting V3$ by varying reconstruction and correctionmethods for the same original camera data acquisitions.

Our aim in this development was to duplicate thepreviously completed 123I-b-CIT striatal V3$ manual anal-ysis for this clinical cohort (15). Therefore, the optimalparameters (N, s, and d) for the OSA program were selectedso as to most closely correlate with the manual analysisobtained by our image-processing researcher. However,these optimized parameters may not be the best for apply-ing OSA to different clinical cohorts with more severe lossof DAT or for monitoring the change in DAT loss resultingfrom disease progression. For example, optimizing theparameters for a group of healthy subjects should probablyyield the smallest group change in V3$ for repeated studiesover several weeks. By a similar argument, optimal param-eters applied to a group of advanced-PD patients wouldbe expected to yield the largest group change in V3$ forrepeated studies over several months. Theoretically, the op-timal parameters for applying OSA would simultaneouslyoptimize the results for both these example groups. OSA is a

flexible tool that can be customized to measure DAT loss atPD diagnosis or the change in DAT with PD progression.

Optimal parameters for OSA may also depend on tech-nical issues such as camera and collimator type, imple-mentation of attenuation or scatter correction, and selectionof a postreconstruction smoothing filter that may meritoptimizing the N, s, and d parameters for a different set ofsite-specific imaging and reconstruction protocols. How-ever, as is the case for several other packaged analysisprograms (17,19–21) that report the number of analysisslices and placement of ROIs, one often assumes thattechnical issues concerning the acquisition and reconstruc-tion of image data can be considered ‘‘acceptably similar’’and, thus, not require optimization for each camera. Thisassumption is reasonable, because optimization of theparameters is a daunting, time-consuming task; the result-ing optimized parameters would perhaps be similar (if notidentical) to the original ones; and camera manufacturershave evolved toward camera and reconstruction character-istics that in most cases are similar and reproducible.

Other investigators have focused on brain image–processingpackages (17–21,26). In its application, OSA most closelyresembles the automated analysis package reported byKoch et al (20,21). Although both software analysis pack-ages ultimately focus on calculating V3$ as described inEquation 1, the 2 methods differ in their philosophies onimage manipulation and ROI-based uptake measurements.

When repositioning each subject’s reconstructed SPECTbrain volume into a standard orientation, OSA applies onlyrigid-body translation and rotation. Independent x-, y-, andz-scaling is applied only to the average scan for optimizingthe subject’s registration accuracy to this standard orienta-tion. No nonlinear warping is applied to the subject’s data,because so doing may add interpolation errors. When SPECTscans are repeated on the same subject, subsequent scans areregistered to the same subject’s first scan (already registeredto the average scan) instead of to the average scan. We believethis method is beneficial, especially when changes in V3$ arecalculated in following disease progression.

With regard to the average scan used for reorienting eachsubject’s SPECT scan, 26 random scans were intentionallyselected so as to yield an ‘‘averaged’’ representation of thelocation of the striata within an ‘‘average’’ subject whowould be imaged in our institute. Given such an averagedgeometry, the intent of the registration process is to removepitch, roll, and yaw positioning differences between sub-jects; that is, the registration process is meant to orient eachsubject reproducibly and standardly within the imagingcamera. Furthermore, we emphasize that although uptakein the striata is more intense, the remaining volume of thebrain (demonstrating nonspecific uptake) plays a larger rolein determining the final orientation of a subject’s brainscan. Because the calculation of best fit relies on minimizingvariances between the average scan and the subject’s scan,the smaller number of voxels in the striata diminishes theimportance of these structures for the best-fit calculation, and

862 THE JOURNAL OF NUCLEAR MEDICINE • Vol. 48 • No. 6 • June 2007

the larger volume of the remaining gray matter plays a largerrole in optimizing the registration. These factors allow us touse a single, stylized average scan for registering either rela-tively disease-free subjects or subjects exhibiting advanced,progressed disease. Additionally, this registration process isnot associated with any localization of the striatal structures.The localization of striata is a totally independent process.

Another difference from other automated approaches isthat OSA does not apply a rigid array of predefined ROItemplates to the internal substructures of the brain. Instead,for the caudate and putamen, the most intense pixel in thecaudate is located and the neighborhood of this location isused for evaluating caudate uptake. We have found thateither for relatively minimal or for severely progressed PD,the most intense area of tracer uptake closely correspondsto the center of the caudate. The placement of the ROI forthe putamen is similarly found by searching at a fixed radialdistance from the caudate until the ROI yields a maximumintensity. The scaled occipital ROI is placed at a consistentdistance from the edge of the posterior brain. These ROIsdo not assume a predefined volume for each structure and,in the case of the caudate and putamen, measure themaximum uptake of the structure. For cardiac applications(27,28), a method of measuring the hottest pixel withradiotracer techniques has been shown to be a reproduc-ible and accurate way to evaluate physiologic functionand is less prone to partial-volume effects resulting fromthe complex spatial geometry of the striatum or atrophicchanges.

Although the most precise results in the comparison ofOSA to manual analysis delivers an R value of 0.97, not allthe variability should be attributed to the automated methodof OSA for placing ROIs. In the case of manual analysis,projection data were reconstructed, attenuation-uncorrectedtransverse slices containing the striata were summed, andthen an ellipse was placed onto the summed slice to applythe Chang correction for attenuation correction using ourimage-processing platform. In the case of OSA analysis,projection data were reconstructed with attenuation correc-tion (using the reconstruction package of the camera manu-facturer) and then made available for subsequent OSAanalysis. Hence, the data used for manual and OSA analysiswere not attenuation-corrected using the same ellipse place-ment. This discrepancy in ellipse placement rests on theoperational characteristics of the workstation and cameracomputer used for reconstruction and analysis. As we havedocumented previously (29), when compared with het-erogeneous attenuation corrections, placement of the atten-uation correction ellipse for homogeneous attenuationcorrection can introduce 5%–10% error in internal brainstructures and, therefore, can degrade the correlation coeffi-cient between the OSA and manual analysis presented here.

Although accuracy and precision were independentlyoptimized in this analysis, clearly, the best method forreporting the results for the 101 subjects studied here wouldbe to apply the parameters yielding optimal precision (i.e.,

r 5 0.9658 with linear regression line y 5 0.7413x 1 0.2692)and then to apply the linear regression as a correction toobtain both the most precise and the most accurate V3$values.

To evaluate the test/retest manual analysis of V3$ in ourinstitute, 2 of our image-processing technologists reana-lyzed 20 randomly selected scans. Although the identicaloriginal SPECT dataset was used as the starting point, theSD in reevaluating the V3$ for these scans showed a 6%variability. This reflects the subjectivity involved in manualanalysis of these scans. OSA compared favorably: Rean-alysis of the same original dataset yielded a variability of0%.

DAT imaging has been used in several clinical studies tomonitor disease progression and assess possible neuro-protective drugs (30,31). OSA would be an extraordinarilyvaluable tool for these studies, enabling objective analysisof imaging data potentially acquired at multiple sites. Infurther studies, we plan to validate OSA in existing clinicalcohorts of PD progression to fully establish the optimalOSA parameters for both PD diagnosis and PD monitoring.

Finally, we believe our development of this automatedimage-processing technique contributes to a recent trendwithin nuclear medicine—a trend toward removing subjec-tivity from the analysis and interpretation of image data.Software packages for the automated processing and anal-ysis of cardiac function images first appeared in the 1990s(32–34). More recently, automated techniques have beenapplied to brain images (17,19–21). A software package forautomated renal function has also been recently reported(35). We believe these developments enhance the clinicalpower of imaging studies in that they reduce the variabilityof quantitative results obtained between image-processingtechnologists and image-acquisition sites. Reduced variabil-ity helps to reduce the number of subjects needed to inves-tigate new diagnostic and therapeutic agents and deliversmore consistent and reproducible clinical evaluations.

CONCLUSION

Our optimized OSA delivers V3$ evaluations thatclosely correlate with a similar evaluation manually appliedby a highly trained image-processing technologist.

REFERENCES

1. Seibyl JP. Imaging studies in movement disorders. Semin Nucl Med. 2003;33:

105–113.

2. Schapira AH, Olanow CW. Neuroprotection in Parkinson disease: mysteries,

myths, and misconceptions. JAMA. 2004;291:358–364.

3. Shoulson I. Experimental therapeutic of neurodegenerative disorders: unmet

needs. Science. 1998;282:1072–1074.

4. Hughes AJ, Daniel SE, Ben-ShlomoY, Lees AJ. The accuracy of diagnosis of

parkinsonian syndromes in a specialist movement disorder service. Brain.

2002;125:861–870.

5. Meara J, Bhowmick BK, Hobson P. Accuracy of diagnosis in patients with

presumed Parkinson’s disease. Age Ageing. 1999;28:99–102.

6. Quinn N. Accuracy of clinical diagnosis in early Parkinson disease. Arch Neurol.

2001;58:316–317.

AUTOMATED PARKINSON’S ANALYSIS • Zubal et al. 863

7. Bennett DA, Beckett LA, Murray AM, et al. Prevalence of parkinsonian signs

and associated mortality in a community population of older people. N Engl J

Med. 1996;334:71–76.

8. Richards M, Stern Y, Mayeux R. Subtle extrapyramidal signs can predict the

development of dementia in elderly individuals. Neurology. 1993;43:2184–2188.

9. Langston JW. The Parkinson’s complex: parkinsonism is just the tip of the

iceberg. Ann Neurol. 2006;59:591–596.

10. Booij J, Habraken JB, Bergmans P, et al. Imaging of dopamine transporters with

iodine-123-FP-CIT SPECT in healthy controls and patients with Parkinson’s

disease. J Nucl Med. 1998;39:1879–1884.

11. Fischman AJ, Bonab AA, Babich JW, et al. [(11)C, (127)I] Altropane: a highly

selective ligand for PET imaging of dopamine transporter sites. Synapse. 2001;

39:332–342.

12. Mozley PD, Schneider JS, Acton PD, et al. Binding of [99mTc]TRODAT-1 to

dopamine transporters in patients with Parkinson’s disease and in healthy

volunteers. J Nucl Med. 2000;41:584–589.

13. Parkinson Study Group. A multicenter assessment of dopamine transporter

imaging with Dopascan�/SPECT in parkinsonism. Neurology. 2000;55:1540–

1547.

14. Seibyl JP, Marek KL, Quinlan D, et al. Decreased single-photon emission

computed tomographic [123I]beta-CIT striatal uptake correlates with symptom

severity in Parkinson’s disease. Ann Neurol. 1995;38:589–598.

15. Jennings DL, Seibyl JP, Oakes D, et al. [123I]b-CIT and single-photon emission

computed tomographic imaging vs clinical evaluation in Parkinsonian syndrome:

unmasking an early diagnosis. Arch Neurol. 2004;61:1224–1229.

16. Catafau AM, Tolosa E. Impact of dopamine transporter SPECT using 123I-

Ioflupane on diagnosis and management of patients with clinically uncertain

Parkinsonian syndromes. Mov Disord. 2004;19:1175–1182.

17. Buchert R, Berding G, Wilke F, et al. IBZM tool: a fully automated expert

system for the evaluation of IBZM SPECT studies. Eur J Nucl Med Mol

Imaging. 2006;33:1073–1083.

18. Morton RJ, Guy MJ, Clauss R, Hinton PJ, Marshall CA, Clarke EA. Comparison

of different methods of DatSCAN quantification. Nucl Med Commun. 2005;26:

1139–1146.

19. Radau PE, Linke R, Slomka PJ, Tatsch K. Optimization of automated

quantification of 123I-IBZM uptake in the striatum applied to parkinsonism.

J Nucl Med. 2000;41:220–227.

20. Koch W, Hamann C, Welsch J, et al. Is iterative reconstruction an alternative to

filtered backprojection in routine processing of dopamine transporter SPECT

studies? J Nucl Med. 2005;46:1804–1811.

21. Koch W, Radau P, Hamann C, Tatsch K. Clinical testing of an optimized

software solution for an automated, observer-independent evaluation of dopa-

mine transporter SPECT studies. J Nucl Med. 2005;46:1109–1118.

22. Chang L. A method for attenuation correction in computed tomography. IEEE

Trans Nucl Sci. 1987;NS-25:638–643.

23. Woods RP, Cherry SR, Mazziotta JC. Rapid automated algorithm for aligning

and reslicing PET images. J Comput Assist Tomogr. 1992;16:620–633.

24. DeKosky S, Marek K. Looking backward to move forward: early detection of

neurodegenerative disorders. Science. 2003;302:830–834.

25. Stern MB. The preclinical detection of Parkinson’s disease: ready for prime

time? Ann Neurol. 2004;56:169–171.

26. Ishii K, Kono AK, Sasaki H, et al. Fully automatic diagnostic system for early-

and late-onset mild Alzheimer’s disease using FDG PET and 3D-SSP. Eur J Nucl

Med Mol Imaging. 2006;33:575–583.

27. Hitzel A, Manrique A, Cribier A, Vera P. Diagnostic value of Tl-201 lung uptake

is dependent on measurement method. J Nucl Cardiol. 2001;8:332–338.

28. Morel O, Pezard P, Furber A, et al. Thallium-201 right lung/heart ratio during

exercise in patients with coronary artery disease: relation to thallium-201

myocardial single-photon emission tomography, rest and exercise left ventricular

function and coronary angiography. Eur J Nucl Med. 1999;26:640–646.

29. Rajeevan N, Zubal IG, Ramsby SQ, Zoghbi SS, Seibyl J, Innis RB. Significance

of nonuniform attenuation correction in quantitative brain SPECT imaging.

J Nucl Med. 1998;39:1719–1726.

30. Parkinson Study Group. Dopamine transporter brain imaging to assess the effects

of Pramipexole vs levodopa Parkinson disease progression. JAMA. 2002;287:

1653–1661.

31. Parkinson Study Group. Levodopa and the progression of Parkinson disease.

N Engl J Med. 2004;351:18–28.

32. Faber TL, Cooke CD, Folks RD, et al. Left ventricular function and perfusion

from gated SPECT perfusion images: an integrated method. J Nucl Med. 1999;

40:650–659.

33. Germano G, Kiat H, Kavanagh PB, et al. Automatic quantification of ejection

fraction from gated myocardial perfusion SPECT. J Nucl Med. 1995;36:2138–2147.

34. Liu YH, Sinusas AJ, Khaimov D, Gebuza BI, Wackers FJ. New hybrid count-

and geometry-based method for quantification of left ventricular volumes and

ejection fraction from ECG-gated SPECT: methodology and validation. J Nucl

Cardiol. 2005;12:55–65.

35. Garcia EV, Taylor A, Halkar R, et al. RENEX: an expert system for the

interpretation of 99mTc-MAG3 scans to detect renal obstruction. J Nucl Med.

2006;47:320–329.

864 THE JOURNAL OF NUCLEAR MEDICINE • Vol. 48 • No. 6 • June 2007